Prediction of acute myocardial infarction or death in acute chest pain patients with machine learning models or first troponin T alone. (14th October 2021)

• Record Type: Journal Article
• Authors: Olsson De Capretz, P; Bjorkelund, A; Mokhtari, A; Bjork, J; Ohlsson, M; Ekelund, U
• Is Part Of: European heart journal. Volume 42(2021)Supplement 1
• Publication Date: 2021-10-14
• Subjects: Artificial Intelligence (Machine Learning, Deep Learning); Cardiology -- Periodicals; Heart -- Diseases -- Periodicals; 616.12005
• DOI: 10.1093/eurheartj/ehab724.3066
• Languages: English
• ISSNs: 0195-668X

Abstract

Abstract: Background: Machine learning approaches are increasingly being explored for use in healthcare systems, but there is a trade-off between increased accuracy and decreased explainability with more complex models. We aimed to evaluate the diagnostic performance for acute myocardial infarction (AMI) or death within 30 days of index visit. Machine learning models were trained using demographic factors, ECG and blood markers, and to compare them to a single high sensitivity TnT (hs-cTnT) value. Methods: Using records from 9519 ED patients from two hospitals in Skåne, Sweden, we created machine learning models based on both logistic regression and artificial neural networks. Inputs in the models varied and included sex and age, first hs-cTnT value at the ED, glucose, creatinine, hemoglobin, and ECG signal data. The models were adapted to meet the following criteria for safe rule-out of 30-day myocardial infarction or death: Negative predictive value (NPV) >99.5% and sensitivity >99%. For rule-in of myocardial infarction or death, a positive predictive value (PPV) of >70% was set. The models were then compared to the performance of a first hs-cTnT <5 ng/L for rule-out, and >51 ng/L for rule-in. The patient population was split by arrival date and models were trained on the initial 50% of patients. Thresholds were selected from the subsequent 25%, and tests were performed on the final 25% (2379 patients). Results: The best model, a convolutional neural network, identified 1309 (55%) patients for rule-out and 125 (5.3%) for rule-in, with the required NPV, sensitivity and PPV. In comparison, a single hs-cTnT value identified 1123 (47.2%) patients for rule-out and 158 (6.6%) for rule-in, but failed to reach the required sensitivity and PPV levels. Conclusions: These results indicate that more complex models are able to safely identify a large proportion of patients for early rule-out or rule-in without the need for serial troponin tests. In future studies attempts should be made to improve the explainability of these models. Funding Acknowledgement: Type of funding sources: Public grant(s) – National budget only. Main funding source(s): VR; grant no. 2019-00198